It's a common task to evaluate a qual and reset the corresponding
expression context. Currently that requires storing the result of the
qual eval, resetting the context, and then reacting on the result. As
that's awkward several places only reset the context next time through
a node. That's not great, so introduce a helper that evaluates and
resets.
It's a bit ugly that it currently uses MemoryContextReset() instead of
ResetExprContext(), but that seems easier than reordering all of
executor.h.
Author: Andres Freund
Discussion: https://postgr.es/m/20180109222544.f7loxrunqh3xjl5f@alap3.anarazel.de
Correct ExecParallelHashTuplePrealloc's estimate of whether the
space_allowed limit is exceeded. Be more consistent about tuples that
are exactly HASH_CHUNK_THRESHOLD in size (they're "small", not "large").
Neither of these things explain the current buildfarm unhappiness, but
they're still bugs.
Thomas Munro, per gripe by me
Discussion: https://postgr.es/m/CAEepm=34PDuR69kfYVhmZPgMdy8pSA-MYbpesEN1SR+2oj3Y+w@mail.gmail.com
The previous coding relied (without any documentation) on the data[]
member of HashMemoryChunkData being at a MAXALIGN'ed offset. If it
was not, the tuples would not be maxaligned either, leading to failures
on alignment-picky machines. While there seems to be no live bug on any
platform we support, this is clearly pretty fragile: any addition to or
rearrangement of the fields in HashMemoryChunkData could break it.
Let's remove the hazard by getting rid of the data[] member and instead
using pointer arithmetic with an explicitly maxalign'ed offset.
Discussion: https://postgr.es/m/14483.1514938129@sss.pgh.pa.us
In a race case, EXPLAIN ANALYZE could fail to display correct nbatch
and size information. Refactor so that participants report only on
batches they worked on rather than trying to report on all of them,
and teach explain.c to consider the HashInstrumentation object from
all participants instead of picking the first one it can find. This
should fix an occasional build farm failure in the "join" regression
test.
Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/30219.1514428346%40sss.pgh.pa.us
When a backend runs out of inner tuples to hash, it should detach from
grow_batch_barrier only after it has flushed all batches to disk and
merged counters, not before. Otherwise a concurrent backend in
ExecParallelHashIncreaseNumBatches() could stop waiting for this
backend and try to read tuples before they have been written. This
commit reorders those operations and should fix the assertion failures
seen occasionally on the build farm since commit
1804284042.
Author: Thomas Munro
Discussion: https://postgr.es/m/E1eRwXy-0004IK-TO%40gemulon.postgresql.org
Introduce parallel-aware hash joins that appear in EXPLAIN plans as Parallel
Hash Join with Parallel Hash. While hash joins could already appear in
parallel queries, they were previously always parallel-oblivious and had a
partial subplan only on the outer side, meaning that the work of the inner
subplan was duplicated in every worker.
After this commit, the planner will consider using a partial subplan on the
inner side too, using the Parallel Hash node to divide the work over the
available CPU cores and combine its results in shared memory. If the join
needs to be split into multiple batches in order to respect work_mem, then
workers process different batches as much as possible and then work together
on the remaining batches.
The advantages of a parallel-aware hash join over a parallel-oblivious hash
join used in a parallel query are that it:
* avoids wasting memory on duplicated hash tables
* avoids wasting disk space on duplicated batch files
* divides the work of building the hash table over the CPUs
One disadvantage is that there is some communication between the participating
CPUs which might outweigh the benefits of parallelism in the case of small
hash tables. This is avoided by the planner's existing reluctance to supply
partial plans for small scans, but it may be necessary to estimate
synchronization costs in future if that situation changes. Another is that
outer batch 0 must be written to disk if multiple batches are required.
A potential future advantage of parallel-aware hash joins is that right and
full outer joins could be supported, since there is a single set of matched
bits for each hashtable, but that is not yet implemented.
A new GUC enable_parallel_hash is defined to control the feature, defaulting
to on.
Author: Thomas Munro
Reviewed-By: Andres Freund, Robert Haas
Tested-By: Rafia Sabih, Prabhat Sahu
Discussion:
https://postgr.es/m/CAEepm=2W=cOkiZxcg6qiFQP-dHUe09aqTrEMM7yJDrHMhDv_RA@mail.gmail.comhttps://postgr.es/m/CAEepm=37HKyJ4U6XOLi=JgfSHM3o6B-GaeO-6hkOmneTDkH+Uw@mail.gmail.com
When a Gather or Gather Merge node is started and stopped multiple
times, accumulate instrumentation data only once, at the end, instead
of after each execution, to avoid recording inflated totals.
Commit 778e78ae9f, the previous attempt
at a fix, instead reset the state after every execution, which worked
for the general instrumentation data but had problems for the additional
instrumentation specific to Sort and Hash nodes.
Report by hubert depesz lubaczewski. Analysis and fix by Amit Kapila,
following a design proposal from Thomas Munro, with a comment tweak
by me.
Discussion: http://postgr.es/m/20171127175631.GA405@depesz.com
If a hash join appears in a parallel query, there may be no hash table
available for explain.c to inspect even though a hash table may have
been built in other processes. This could happen either because
parallel_leader_participation was set to off or because the leader
happened to hit the end of the outer relation immediately (even though
the complete relation is not empty) and decided not to build the hash
table.
Commit bf11e7ee introduced a way for workers to exchange
instrumentation via the DSM segment for Sort nodes even though they
are not parallel-aware. This commit does the same for Hash nodes, so
that explain.c has a way to find instrumentation data from an
arbitrary participant that actually built the hash table.
Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm%3D3DUQC2-z252N55eOcZBer6DPdM%3DFzrxH9dZc5vYLsjaA%40mail.gmail.com
The lower case spellings are C and C++ standard and are used in most
parts of the PostgreSQL sources. The upper case spellings are only used
in some files/modules. So standardize on the standard spellings.
The APIs for ICU, Perl, and Windows define their own TRUE and FALSE, so
those are left as is when using those APIs.
In code comments, we use the lower-case spelling for the C concepts and
keep the upper-case spelling for the SQL concepts.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
This allows us to add stack-depth checks the first time an executor
node is called, and skip that overhead on following
calls. Additionally it yields a nice speedup.
While it'd probably have been a good idea to have that check all
along, it has become more important after the new expression
evaluation framework in b8d7f053c5 - there's no stack depth
check in common paths anymore now. We previously relied on
ExecEvalExpr() being executed somewhere.
We should move towards that model for further routines, but as this is
required for v10, it seems better to only do the necessary (which
already is quite large).
Author: Andres Freund, Tom Lane
Reported-By: Julien Rouhaud
Discussion:
https://postgr.es/m/22833.1490390175@sss.pgh.pa.ushttps://postgr.es/m/b0af9eaa-130c-60d0-9e4e-7a135b1e0c76@dalibo.com
In a followup commit ExecProcNode(), and especially the large switch
it contains, will largely be replaced by a function pointer directly
to the correct node. The node functions will then get invoked by a
thin inline function wrapper. To avoid having to include miscadmin.h
in headers - CHECK_FOR_INTERRUPTS() - move the interrupt checks into
the individual executor routines.
While looking through all executor nodes, I noticed a number of
arguably missing interrupt checks, add these too.
Author: Andres Freund, Tom Lane
Reviewed-By: Tom Lane
Discussion:
https://postgr.es/m/22833.1490390175@sss.pgh.pa.us
Don't move parenthesized lines to the left, even if that means they
flow past the right margin.
By default, BSD indent lines up statement continuation lines that are
within parentheses so that they start just to the right of the preceding
left parenthesis. However, traditionally, if that resulted in the
continuation line extending to the right of the desired right margin,
then indent would push it left just far enough to not overrun the margin,
if it could do so without making the continuation line start to the left of
the current statement indent. That makes for a weird mix of indentations
unless one has been completely rigid about never violating the 80-column
limit.
This behavior has been pretty universally panned by Postgres developers.
Hence, disable it with indent's new -lpl switch, so that parenthesized
lines are always lined up with the preceding left paren.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
Change pg_bsd_indent to follow upstream rules for placement of comments
to the right of code, and remove pgindent hack that caused comments
following #endif to not obey the general rule.
Commit e3860ffa4d wasn't actually using
the published version of pg_bsd_indent, but a hacked-up version that
tried to minimize the amount of movement of comments to the right of
code. The situation of interest is where such a comment has to be
moved to the right of its default placement at column 33 because there's
code there. BSD indent has always moved right in units of tab stops
in such cases --- but in the previous incarnation, indent was working
in 8-space tab stops, while now it knows we use 4-space tabs. So the
net result is that in about half the cases, such comments are placed
one tab stop left of before. This is better all around: it leaves
more room on the line for comment text, and it means that in such
cases the comment uniformly starts at the next 4-space tab stop after
the code, rather than sometimes one and sometimes two tabs after.
Also, ensure that comments following #endif are indented the same
as comments following other preprocessor commands such as #else.
That inconsistency turns out to have been self-inflicted damage
from a poorly-thought-through post-indent "fixup" in pgindent.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
The mess cleaned up in commit da0759600 is clear evidence that it's a
bug hazard to expect the caller of get_attstatsslot()/free_attstatsslot()
to provide the correct type OID for the array elements in the slot.
Moreover, we weren't even getting any performance benefit from that,
since get_attstatsslot() was extracting the real type OID from the array
anyway. So we ought to get rid of that requirement; indeed, it would
make more sense for get_attstatsslot() to pass back the type OID it found,
in case the caller isn't sure what to expect, which is likely in binary-
compatible-operator cases.
Another problem with the current implementation is that if the stats array
element type is pass-by-reference, we incur a palloc/memcpy/pfree cycle
for each element. That seemed acceptable when the code was written because
we were targeting O(10) array sizes --- but these days, stats arrays are
almost always bigger than that, sometimes much bigger. We can save a
significant number of cycles by doing one palloc/memcpy/pfree of the whole
array. Indeed, in the now-probably-common case where the array is toasted,
that happens anyway so this method is basically free. (Note: although the
catcache code will inline any out-of-line toasted values, it doesn't
decompress them. At the other end of the size range, it doesn't expand
short-header datums either. In either case, DatumGetArrayTypeP would have
to make a copy. We do end up using an extra array copy step if the element
type is pass-by-value and the array length is neither small enough for a
short header nor large enough to have suffered compression. But that
seems like a very acceptable price for winning in pass-by-ref cases.)
Hence, redesign to take these insights into account. While at it,
convert to an API in which we fill a struct rather than passing a bunch
of pointers to individual output arguments. That will make it less
painful if we ever want further expansion of what get_attstatsslot can
pass back.
It's certainly arguable that this is new development and not something to
push post-feature-freeze. However, I view it as primarily bug-proofing
and therefore something that's better to have sooner not later. Since
we aren't quite at beta phase yet, let's put it in.
Discussion: https://postgr.es/m/16364.1494520862@sss.pgh.pa.us
This replaces the old, recursive tree-walk based evaluation, with
non-recursive, opcode dispatch based, expression evaluation.
Projection is now implemented as part of expression evaluation.
This both leads to significant performance improvements, and makes
future just-in-time compilation of expressions easier.
The speed gains primarily come from:
- non-recursive implementation reduces stack usage / overhead
- simple sub-expressions are implemented with a single jump, without
function calls
- sharing some state between different sub-expressions
- reduced amount of indirect/hard to predict memory accesses by laying
out operation metadata sequentially; including the avoidance of
nearly all of the previously used linked lists
- more code has been moved to expression initialization, avoiding
constant re-checks at evaluation time
Future just-in-time compilation (JIT) has become easier, as
demonstrated by released patches intended to be merged in a later
release, for primarily two reasons: Firstly, due to a stricter split
between expression initialization and evaluation, less code has to be
handled by the JIT. Secondly, due to the non-recursive nature of the
generated "instructions", less performance-critical code-paths can
easily be shared between interpreted and compiled evaluation.
The new framework allows for significant future optimizations. E.g.:
- basic infrastructure for to later reduce the per executor-startup
overhead of expression evaluation, by caching state in prepared
statements. That'd be helpful in OLTPish scenarios where
initialization overhead is measurable.
- optimizing the generated "code". A number of proposals for potential
work has already been made.
- optimizing the interpreter. Similarly a number of proposals have
been made here too.
The move of logic into the expression initialization step leads to some
backward-incompatible changes:
- Function permission checks are now done during expression
initialization, whereas previously they were done during
execution. In edge cases this can lead to errors being raised that
previously wouldn't have been, e.g. a NULL array being coerced to a
different array type previously didn't perform checks.
- The set of domain constraints to be checked, is now evaluated once
during expression initialization, previously it was re-built
every time a domain check was evaluated. For normal queries this
doesn't change much, but e.g. for plpgsql functions, which caches
ExprStates, the old set could stick around longer. The behavior
around might still change.
Author: Andres Freund, with significant changes by Tom Lane,
changes by Heikki Linnakangas
Reviewed-By: Tom Lane, Heikki Linnakangas
Discussion: https://postgr.es/m/20161206034955.bh33paeralxbtluv@alap3.anarazel.de
The loops in ExecHashJoinNewBatch(), ExecHashIncreaseNumBatches(), and
ExecHashRemoveNextSkewBucket() are all capable of iterating over many
tuples without ever doing a CHECK_FOR_INTERRUPTS, so that the backend
might fail to respond to SIGINT or SIGTERM for an unreasonably long time.
Fix that. In the case of ExecHashJoinNewBatch(), it seems useful to put
the added CHECK_FOR_INTERRUPTS into ExecHashJoinGetSavedTuple() rather
than directly in the loop, because that will also ensure that both
principal code paths through ExecHashJoinOuterGetTuple() will do a
CHECK_FOR_INTERRUPTS, which seems like a good idea to avoid surprises.
Back-patch to all supported branches.
Tom Lane and Thomas Munro
Discussion: https://postgr.es/m/6044.1487121720@sss.pgh.pa.us
Since 69f4b9c plain expression evaluation (and thus normal projection)
can't return sets of tuples anymore. Thus remove code dealing with
that possibility.
This will require adjustments in external code using
ExecEvalExpr()/ExecProject() - that should neither be hard nor very
common.
Author: Andres Freund and Tom Lane
Discussion: https://postgr.es/m/20160822214023.aaxz5l4igypowyri@alap3.anarazel.de
I found that half a dozen (nearly 5%) of our AllocSetContextCreate calls
had typos in the context-sizing parameters. While none of these led to
especially significant problems, they did create minor inefficiencies,
and it's now clear that expecting people to copy-and-paste those calls
accurately is not a great idea. Let's reduce the risk of future errors
by introducing single macros that encapsulate the common use-cases.
Three such macros are enough to cover all but two special-purpose contexts;
those two calls can be left as-is, I think.
While this patch doesn't in itself improve matters for third-party
extensions, it doesn't break anything for them either, and they can
gradually adopt the simplified notation over time.
In passing, change TopMemoryContext to use the default allocation
parameters. Formerly it could only be extended 8K at a time. That was
probably reasonable when this code was written; but nowadays we create
many more contexts than we did then, so that it's not unusual to have a
couple hundred K in TopMemoryContext, even without considering various
dubious code that sticks other things there. There seems no good reason
not to let it use growing blocks like most other contexts.
Back-patch to 9.6, mostly because that's still close enough to HEAD that
it's easy to do so, and keeping the branches in sync can be expected to
avoid some future back-patching pain. The bugs fixed by these changes
don't seem to be significant enough to justify fixing them further back.
Discussion: <21072.1472321324@sss.pgh.pa.us>
Commit 45f6240a8f added an assumption in ExecHashIncreaseNumBatches
and ExecHashIncreaseNumBuckets that they could find all tuples in the main
hash table by iterating over the "dense storage" introduced by that patch.
However, ExecHashRemoveNextSkewBucket continued its old practice of simply
re-linking deleted skew tuples into the main table's hashchains. Hence,
such tuples got lost during any subsequent increase in nbatch or nbuckets,
and would never get joined, as reported in bug #13908 from Seth P.
I (tgl) think that the aforesaid commit has got multiple design issues
and should be reworked rather completely; but there is no time for that
right now, so band-aid the problem by making ExecHashRemoveNextSkewBucket
physically copy deleted skew tuples into the "dense storage" arena.
The added test case is able to exhibit the problem by means of fooling the
planner with a WHERE condition that it will underestimate the selectivity
of, causing the initial nbatch estimate to be too small.
Tomas Vondra and Tom Lane. Thanks to David Johnston for initial
investigation into the bug report.
Commit 30d7ae3c76 introduced an HJDEBUG
stanza that probably didn't compile at the time, and definitely doesn't
compile now, because it refers to a nonexistent variable. It doesn't seem
terribly useful anyway, so just get rid of it.
While I'm fooling with it, use %z modifier instead of the obsolete hack of
casting size_t to unsigned long, and include the HashJoinTable's address in
each printout so that it's possible to distinguish the activities of
multiple hashjoins occurring in one query.
Noted while trying to use HJDEBUG to investigate bug #13908. Back-patch
to 9.5, because code that doesn't compile is certainly not very helpful.
On reflection, the submitted patch didn't really work to prevent the
request size from exceeding MaxAllocSize, because of the fact that we'd
happily round nbuckets up to the next power of 2 after we'd limited it to
max_pointers. The simplest way to enforce the limit correctly is to
round max_pointers down to a power of 2 when it isn't one already.
(Note that the constraint to INT_MAX / 2, if it were doing anything useful
at all, is properly applied after that.)
Limit the size of the hashtable pointer array to not more than
MaxAllocSize, per reports from Kouhei Kaigai and others of "invalid memory
alloc request size" failures. There was discussion of allowing the array
to get larger than that by using the "huge" palloc API, but so far no proof
that that is actually a good idea, and at this point in the 9.5 cycle major
changes from old behavior don't seem like the way to go.
Fix a rather serious secondary bug in the new code, which was that it
didn't ensure nbuckets remained a power of 2 when recomputing it for the
multiple-batch case.
Clean up sloppy division of labor between ExecHashIncreaseNumBuckets and
its sole call site.
Commit 8cce08f168 used a left-shift
on a literal of 1 that could (in large allocations) be shifted by
31 or more bits. This was assigned to a local variable that was
already declared to be a long to protect against overruns of int,
but the literal in this shift needs to be declared long to allow it
to work correctly in some compilers.
Backpatch to 9.5, where the bug was introduced.
Report and patch by KaiGai Kohei, slighly modified based on
discussion.
Although I think on all modern machines floating division by zero
results in Infinity not SIGFPE, we still don't want infinities
running around in the planner's costing estimates; too much risk
of that leading to insane behavior.
grouping_planner() failed to consider the possibility that final_rel
might be known dummy and hence have zero rowcount. (I wonder if it
would be better to set a rows estimate of 1 for dummy relations?
But at least in the back branches, changing this convention seems
like a bad idea, so I'll leave that for another day.)
Make certain that get_variable_numdistinct() produces a nonzero result.
The case that can be shown to be broken is with stadistinct < 0.0 and
small ntuples; we did not prevent the result from rounding to zero.
For good luck I applied clamp_row_est() to all the nonconstant return
values.
In ExecChooseHashTableSize(), Assert that we compute positive nbuckets
and nbatch. I know of no reason to think this isn't the case, but it
seems like a good safety check.
Per reports from Piotr Stefaniak. Back-patch to all active branches.
Fix some places where pgindent did silly stuff, often because project
style wasn't followed to begin with. (I've not touched the atomics
headers, though.)
This requires changing quite a few places that were depending on
sizeof(HeapTupleHeaderData), but it seems for the best.
Michael Paquier, some adjustments by me
If we expect batching at the very beginning, we size nbuckets for
"full work_mem" (see how many tuples we can get into work_mem,
while not breaking NTUP_PER_BUCKET threshold).
If we expect to be fine without batching, we start with the 'right'
nbuckets and track the optimal nbuckets as we go (without actually
resizing the hash table). Once we hit work_mem (considering the
optimal nbuckets value), we keep the value.
At the end of the first batch, we check whether (nbuckets !=
nbuckets_optimal) and resize the hash table if needed. Also, we
keep this value for all batches (it's OK because it assumes full
work_mem, and it makes the batchno evaluation trivial). So the
resize happens only once.
There could be cases where it would improve performance to allow
the NTUP_PER_BUCKET threshold to be exceeded to keep everything in
one batch rather than spilling to a second batch, but attempts to
generate such a case have so far been unsuccessful; that issue may
be addressed with a follow-on patch after further investigation.
Tomas Vondra with minor format and comment cleanup by me
Reviewed by Robert Haas, Heikki Linnakangas, and Kevin Grittner
Since this makes the bucket headers use ~10x as much memory, properly
account for that memory when we figure out whether everything fits
in work_mem. This might result in some cases that previously used
only a single batch getting split into multiple batches, but it's
unclear as yet whether we need defenses against that case, and if so,
what the shape of those defenses should be.
It's worth noting that even in these edge cases, users should still be
no worse off than they would have been last week, because commit
45f6240a8f saved a big pile of memory
on exactly the same workloads.
Tomas Vondra, reviewed and somewhat revised by me.
Instead of palloc'ing each HashJoinTuple individually, allocate 32kB chunks
and pack the tuples densely in the chunks. This avoids the AllocChunk
header overhead, and the space wasted by standard allocator's habit of
rounding sizes up to the nearest power of two.
This doesn't contain any planner changes, because the planner's estimate of
memory usage ignores the palloc overhead. Now that the overhead is smaller,
the planner's estimates are in fact more accurate.
Tomas Vondra, reviewed by Robert Haas.
The dynahash code requires the number of buckets in a hash table to fit
in an int; but since we calculate the desired hash table size dynamically,
there are various scenarios where we might calculate too large a value.
The resulting overflow can lead to infinite loops, division-by-zero
crashes, etc. I (tgl) had previously installed some defenses against that
in commit 299d171652, but that covered only one
call path. Moreover it worked by limiting the request size to work_mem,
but in a 64-bit machine it's possible to set work_mem high enough that the
problem appears anyway. So let's fix the problem at the root by installing
limits in the dynahash.c functions themselves.
Trouble report and patch by Jeff Davis.
This reduces unnecessary exposure of other headers through htup.h, which
is very widely included by many files.
I have chosen to move the function prototypes to the new file as well,
because that means htup.h no longer needs to include tupdesc.h. In
itself this doesn't have much effect in indirect inclusion of tupdesc.h
throughout the tree, because it's also required by execnodes.h; but it's
something to explore in the future, and it seemed best to do the htup.h
change now while I'm busy with it.
This commit changes index-only scans so that data is read directly from the
index tuple without first generating a faux heap tuple. The only immediate
benefit is that indexes on system columns (such as OID) can be used in
index-only scans, but this is necessary infrastructure if we are ever to
support index-only scans on expression indexes. The executor is now ready
for that, though the planner still needs substantial work to recognize
the possibility.
To do this, Vars in index-only plan nodes have to refer to index columns
not heap columns. I introduced a new special varno, INDEX_VAR, to mark
such Vars to avoid confusion. (In passing, this commit renames the two
existing special varnos to OUTER_VAR and INNER_VAR.) This allows
ruleutils.c to handle them with logic similar to what we use for subplan
reference Vars.
Since index-only scans are now fundamentally different from regular
indexscans so far as their expression subtrees are concerned, I also chose
to change them to have their own plan node type (and hence, their own
executor source file).
This provides information about the numbers of tuples that were visited
but not returned by table scans, as well as the numbers of join tuples
that were considered and discarded within a join plan node.
There is still some discussion going on about the best way to report counts
for outer-join situations, but I think most of what's in the patch would
not change if we revise that, so I'm going to go ahead and commit it as-is.
Documentation changes to follow (they weren't in the submitted patch
either).
Marko Tiikkaja, reviewed by Marc Cousin, somewhat revised by Tom
This is advantageous first because it allows us to hash the smaller table
regardless of the outer-join type, and second because hash join can be more
flexible than merge join in dealing with arbitrary join quals in a FULL
join. For merge join all the join quals have to be mergejoinable, but hash
join will work so long as there's at least one hashjoinable qual --- the
others can be any condition. (This is true essentially because we don't
keep per-inner-tuple match flags in merge join, while hash join can do so.)
To do this, we need a has-it-been-matched flag for each tuple in the
hashtable, not just one for the current outer tuple. The key idea that
makes this practical is that we can store the match flag in the tuple's
infomask, since there are lots of bits there that are of no interest for a
MinimalTuple. So we aren't increasing the size of the hashtable at all for
the feature.
To write this without turning the hash code into even more of a pile of
spaghetti than it already was, I rewrote ExecHashJoin in a state-machine
style, similar to ExecMergeJoin. Other than that decision, it was pretty
straightforward.
Andres Freund